Generative Adversarial Network with Policy Gradient for Text Summarization

Abstractive text summarization is the task of generating meaningful summary from a given document (short or long). This is a very challenging task for longer documents, since they suffer from repetitions (redundancy) when the given document is long and the generated summary should contain multi-sentences. In this paper we present an approach for applying generative adversarial networks in abstractive text summarization tasks with a novel time-decay attention mechanism. The data generator is modeled as a stochastic policy in reinforcement learning. The generator's goal is to generate summaries which are difficult to be discriminated from real summaries. The discriminator aims to estimate the probability that a summary came from the training data rather than the generator to guide the training of the generative model. This framework corresponds to a minimax two-player game. Qualitatively and quantitatively experimental results (human evaluations and ROUGE scores) show that our model can generate more relevant, less repetitive, grammatically correct, preferable by humans and is promising in solving the abstractive text summarization task.